Authors: Jun Wang, Klaus Mueller
Abstract: Deriving the exact casual model that governs the relations between variables in a multidimensional dataset is difficult in practice. It is because causal inference algorithms by themselves typically cannot encode an adequate amount of domain knowledge to break all ties. To that end, visual analytic approaches are considered a feasible alternative to fully automated methods. However, when applying this visual causality analysis in real-world scenarios many practical issues need to be solved. This paper focuses on these practical aspects of visual causality analysis. The most imperative of these aspects is posed by Simpson’ Paradox. It implies the existence of multiple causal models differing in both structure and parameter depending on how the data is subdivided. We propose a comprehensive interface that engages human experts in identifying these subdivisions and allowing them to establish the corresponding causal models via a rich set of interactive facilities. Other features of our interface include: (1) a new causal network visualization that emphasizes the flows of causal dependencies, (2) a model scoring mechanism with visual hints for interactive model refinement, and (3) flexible approaches for handling heterogeneous \ data. We demonstrate our framework with several real-world datasets from a diverse set of domains.